Appraisal Advisor — Valuation Report Review Agent
Role: Product Manager & Lead Developer
Reviewing a valuation report depends on how alert the reviewer was, so outcomes vary and tedious arithmetic errors slip through. This automation agent receives the report in a conversation, applies a fixed checklist to it, and returns a review in which every finding is tied to the article that supports it — so the output can be checked rather than believed.
A PDF arrives in a conversation, its numeric fields and dates are extracted, and 73 checks across eight sections are applied, resting on five regulatory references: the approved report form, the executive regulation, the financing rules, IVS 2025, and the indicative price guide. Thirty of those items form a conformity matrix marking each one met, partial, or absent. The review returns to the same conversation with a summary, severity counters and an overall verdict, and is archived automatically — into a monthly folder the system creates itself the first time and remembers afterwards, holding a formatted PDF of every review. And twelve of the 73 rules target the recurring error patterns of the report-issuing system itself: the empty deposit code, the all-zero adjustment grid, and their kin.
What sets it apart is that it separates the deterministic from the probabilistic: seven critical numeric checks run in code rather than in a language model — expiry, date ordering, area × price against the stated total within 0.5%, a gap above 15% between two reports, an all-zero adjustment grid — so they return the same answer on every run, and the model inserts them as they are without recomputing. And when comparables are analysed a simple average is forbidden: weighting favours the least-adjusted and most recent, and a divergence gate at 20% stops the weighting altogether and calls for the methods to be reconsidered rather than emitting a misleading figure. Before it hands anything over it checks itself — is every item backed? is each citation in its right place and full form? — and a conversation keeps its memory, so a follow-up question lands in the context of the review it follows. The pipeline is logically complete and exercised through manual runs, not yet deployed, pending a decision to activate — and there is no approval gate inside it; the human barrier sits outside: the output is advisory and sign-off remains with a licensed valuer.
Key Features
Seven Checks in Code, Not in a Model
Critical numbers and dates are checked programmatically, so the result is identical every run and the model inserts it without recomputing.
73 Rules Applied Equally to Every Report
Eight sections, of which 30 items form a conformity matrix marking each one met, partial, or absent.
No Finding Without a Citation
Every finding names its regulation and article; one without a citation is rejected.
A Divergence Gate That Stops the Weighting
When comparables spread beyond 20% it produces no weighted figure and asks for the methods to be reconsidered.
It Knows Its Limit and Stops There
It issues no alternative value, approves nothing and signs nothing; it offers a reasonableness opinion and leaves the decision to a licensed valuer.
It Fails Out Loud
A failed extraction surfaces as an explicit red finding saying the deterministic checks did not run, and a non-PDF file gets a reply rather than silence.
The Business Problem
Reviewing a valuation report is manual: dozens of pages checked line by line against the approved form, the executive regulation, the financing rules and IVS — so the outcome varies with whichever reviewer, and how alert they were. Errors like an empty deposit code, an all-zero adjustment grid, or a valuation date later than the issue date slip through, because catching them is tedious arithmetic rather than expert judgement.
Why We Built It
Fix the standard rather than the reviewer: one checklist applied with the same rigour every time, each finding tied to its article so it can be checked rather than believed. The split is deliberate — what arithmetic can settle stays in code, and what needs judgement goes to a model constrained to cite.
How We Built It
An n8n pipeline that starts from a message carrying a PDF. An extraction model returns the fields as JSON under fixed keys, then a JavaScript check engine applies the seven deterministic rules and passes them through ready-made. The review model applies the checklist, retrieving reference texts verbatim when unsure rather than guessing them, and emits an object under closed value lists — three statuses, three severities, three verdicts — enforced by an output parser. From it an HTML report is built, converted to PDF, and archived.